liveBook · Manning

Memory isolation

As explained in AI Governance

Memory isolation is the separation and restriction of agent memory so that information or instructions in one task or scope cannot improperly affect another.

Memory isolation limits how information persists between tasks, reducing the chance that malicious instructions planted in one session affect later work. Persistent memory improves agent usefulness but also creates an attack surface. section 4.8.7

A tiered design controls retention and write access:

  • Ephemeral working memory resets after each task. section 4.8.7
  • Bounded session logs expire after ninety days. section 4.8.7
  • Permanent knowledge changes require explicit human review; paraphrasing and summarizing before storage can avoid preserving injected instructions verbatim. section 4.8.7

Cross-session memory poisoning

An attacker plants malicious instructions in session memory. Automatic expiration limits how long the poisoned patterns can persist; human review gates permanent updates. section 4.8.7

How AI Governance covers it

How deep it goes
Distinguishes three memory tiers and describes retention, review, and safer persistence practices.
Examples
Uses planted instructions, task resets, ninety-day log expiration, and reviewed permanent updates.
What you can do after
Configure memory retention tiers and gate permanent writes with human review.
How it connects
Connects persistent memory with memory poisoning and sandboxing as complementary containment controls.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of AI Governance. The section links point to the parts of the book it draws on.